Read committed
SeKwonLee opened this issue · 0 comments
Current implementations only ensure the lowest level of isolation (Read Uncommitted) for some read operations such as scan, negative lookup, and verification for value existence, since they are based on normal CASs or temporal stores coupled with cache line flush instructions. However, it is not the fundamental limitation of RECIPE conversions. You can easily extend them, following RECIPE conversions, to guarantee the higher level of isolation (Read Committed) by replacing each final commit stores (such as pointer swap) coupled with cache line flushes with non-temporal stores
coupled with memory fence for lock-based implementations
including P-CLHT, P-HOT, P-ART, and P-Masstree. For lock-free implementations
such as P-Bwtree, you can either add additional flushes only after loads to final commit stores or replace volatile CASs coupled with cache line flush instructions with alternative software-based atomic-persistent primitives such as either Link-and-Persist
(paper, code) or PSwCAS
(paper, code).